[Kernel] Add per-layer LF-to-PF KV transfer for LayerSplit HiCache

[dblate]

Motivation

Enable layer-sharded MLA/DSA HiCache write-back for page-first host storage when the JIT path cannot handle the per-item size. This covers the GLM-5.2 LayerSplit CP path without changing cache quantization behavior.

Modifications

• Add CUDA, MUSA, and ROCm registration for a per-layer LF-to-PF transfer operator.
• Expose the operator through the Python kvcacheio bindings.
• Add BF16/FP16 round-trip coverage for the new transfer path.

Accuracy Tests

• H20, CUDA 12.9: test_kvcacheio.py -k page_first_roundtrip -> 2 passed.
• The test covers BF16 and FP16 round trips through the new operator.

Speed Tests and Profiling

H20 microbenchmark, 8192 items, 656 bytes/item, 100 iterations:

LF-to-PF per-layer: 0.2747 ms, 19.56 GB/s
LF-to-LF per-layer: 0.1892 ms, 28.40 GB/s

LF-to-PF is a layout-transforming fallback, so this is a directional kernel comparison rather than an end-to-end serving throughput result.

Checklist

• Format code with pre-commit.
• Add unit tests according to the contribution guide.
• Update documentation according to the contribution guide.
• Provide GPU accuracy and directional benchmark results.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33379044561
Latest PR Test (Extra): ❌ Run #33379044418
Latest PR Test (AMD ROCm 7.2): ❌ Run #33379044517

[dblate]

/tag-and-rerun-ci

[dblate]

补充 NVIDIA GPU 验证：在 H20（CUDA 12.9）开发空间运行 test_kvcacheio.py -k page_first_roundtrip，结果为 2 passed。核心 LF-to-PF kernel round-trip 通过。

当前 CI 仍未开始真实测试，原因是缺少上游要求的 run-ci 标签；请维护者按贡献指南执行 /tag-and-rerun-ci。

[dblate]

H20 microbenchmark (CUDA 12.9, 8192 items, 656 bytes/item, 100 iterations):

LF-to-PF per-layer: 0.2747 ms, 19.56 GB/s
LF-to-LF per-layer: 0.1892 ms, 28.40 GB/s

The LF-to-PF path is a layout-transforming fallback, so this is a directional kernel comparison rather than an end-to-end serving throughput result.